DTIC ADA458653: Multiple Approaches to Robust Speech Recognition

7 Pages·0.65 MB·English

by  Defense Technical Information Center

MULTIPLE APPROACHES TO ROBUST SPEECH RECOGNITION
Richard M. Stern, Fu-Hua Liu, Yoshiaki Ohshima, Thomas M. Sullivan, Alejandro Acero*
Department of Electrical and Computer Engineering
School of Computer Science
Carnegie Mellon University
Pittsburgh, PA 15213
ABSTRACT our results for the DARPA benchmark evaluation for
robust speech recognition for the ATIS task, discussing the
This paper compares several different approaches to robust effectiveness of our methods of acoustical pre-
speech recognition. We review CMU's ongoing research in the preprocessing in the context of this task. Second, we
use of acoustical pre-proeessing to achieve robust speech recog- describe and compare the effectiveness of three com-
nition, and we present the results of the first evaluation of pre- plementary methods of signal processing for robust speech
processing in the context of the DARPA standard ATIS domain recognition: acoustical pre-processing, microphone array
for spoken language systems. We also describe and compare the processing, and the use of physiologically-motivated
models of peripheral signal processing.
effectiveness of three complementary methods of signal process-
ing for robust speech recognition: acoustical pre-procossing,
microphone array processing, and the use of physiologically-
2. ACOUSTICAL PRE-PROCESSING
motivated models of peripheral signal processing. Recognition
error rates are presented using these three approaches in isolation
We have found that two major factors degrading the per-
and in combination with each other for the speaker-independent
formance of speech recognition systems using desktop
continuous alphanumeric census speech recognition task.
microphones in normal office environments are additive
noise and unknown linear filtering. We showed in 2 that
simultaneous joint compensation for the effects of additive
1. INTRODUCTION
noise and linear filtering is needed to achieve maximal
robustness with respect to acoustical differences between
The need for speech recognition systems and spoken lan- the training and testing environments of a speech recog-
guage systems to be robust with respect to their acoustical nition system. We described in 2 two algorithms that can
environment has become more widely appreciated in perform such joint compensation, based on additive correc-
recent years (e.g. 1). tions to the cepstral coefficients of the speech waveform.
Results of several studies have demonstrated that even The first compensation algorithm, SNR-Dependent
automatic speech recognition systems that are designed to Cepstral Normalization (SDCN), applies an additive cor-
be speaker independent can perform very poorly when they rection in the cepstral domain that depends exclusively on
are tested using a different type of microphone or acous- the instantaneous SNR of the signal. This correction vec-
tical environment from the one with which they were tor equals the average difference in cepstra between simul-
trained (e.g. 2, 3), even in a relatively quiet office en- taneous "stereo" recordings of speech samples from both
vironment. Applications such as speech recognition over the training and testing environments at each SNR of
telephones, in automobiles, on a factory floor, or outdoors speech in the testing environment. At high SNRs, this
demand an even greater degree of environmental robust- correction vector primarily compensates for differences in
ness. spectral flit between the training and testing environments
(in a manner similar to the blind deconvolution procedure
The CMU speech group is committed to the development
first proposed by Stockham et .la 6), while at low SNRs
of speech recognition systems that are robust with respect
the vector provides a form of noise subtraction (in a man-
to environmental variation, just as it has been an early
ner similar to the spectral subtraction algorithm first
proponent of speaker-independent recognition. While most
proposed by Boll 7). The SDCN algorithm is simple and
of our work presented to date has described new acoustical
effective, but it requires environment-specific training.
pre-processing algorithms (e.g. 2, 4, 5, we have always
regarded pre-processing as one of several approaches that The second compensation algorithm, Codeword-
must be developed in concert to achieve robust recog- Dependent Cepstral Normalization (CDCN), uses EM
nition. techniques to compute ML estimates of the parameters
characterizing the contributions of additive noise and
The purpose of this paper is twofold. First, we describe
linear filtering that when applied in inverse fashion to the
tneserP* address: Telef6nica Invesfigaci6n y Desarrollo, Emilio Vargas ,6 Madrid 28043, Spain
274
Report Documentation Page Form Approved
OMB No. 0704-0188
Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington
VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it

does not display a currently valid OMB control number.
1. REPORT DATE 3. DATES COVERED
1992 2. REPORT TYPE 00-00-1992 to 00-00-1992
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
Multiple Approaches to Robust Speech Recognition
5b. GRANT NUMBER
5c. PROGRAM ELEMENT NUMBER
6. AUTHOR(S) 5d. PROJECT NUMBER
5e. TASK NUMBER
5f. WORK UNIT NUMBER
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION
Carnegie Mellon University,School of Computer REPORT NUMBER
Science,Pittsburgh,PA,15213
9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)
11. SPONSOR/MONITOR’S REPORT
NUMBER(S)
12. DISTRIBUTION/AVAILABILITY STATEMENT
Approved for public release; distribution unlimited
13. SUPPLEMENTARY NOTES
14. ABSTRACT
15. SUBJECT TERMS
16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF
ABSTRACT OF PAGES RESPONSIBLE PERSON

a. REPORT b. ABSTRACT c. THIS PAGE 6
unclassified unclassified unclassified
Standard Form 298 (Rev. 8-98)
Prescribed by ANSI Std Z39-18
cepstra of an incoming utterance produce an ensemble of
ALGO- ENVIRN. COM- ERR
cepstral coefficients that best match (in the ML sense) the
RITHM SPEC? PLEXITY RATE
cepstral coefficients of the incoming speech in the testing
environment to the locations of VQ codewords in the train- NONE NO NONE 68.6%
ing environment. The CDCN algorithm has the advantage
SDCN YES MINIMAL 27.6%
that it does not require a priori knowledge of the testing
environment (in the form of stereo training data in the CDCN NO GREATER 24.3%
training and testing environments), but it is much more
computationally demanding than the SDCN algorithm. BSDCN NO MINIMAL 30.0%
Compared to the SDCN algorithm, the CDCN algorithm
uses a greater amount of structural knowledge about the
nature of the degradations to the speech signal in order to Table 1: Comparison of recognition accuracy of XNIHPS
achieve good recognition accuracy. The SDCN algorithm, with no processing and the CDCN, SDCN, and BSDCN
on the other hand, derives its compensation vectors en- algorithms. The system was trained using the CLSTLK
tirely from empirical observations of differences between microphone and tested using the PZM6FS microphone.
atad obtained from the training and testing environments. Training and testing on the CLSTLK produces a recog-
nition accuracy of 86.9%, while training and testing on the
100
PZM6FS produces 76.2%
O.
Train CLSTLK, Tul PZM6FS
6O SNRs in the training and testing environments by use of
% traditional nonlinear warping techniques 8 on histograms
40 "0... of SNRs from each of the two environments 5. Table 1
................... ~ ..... .~..Q
compares the environmental specificity, computational
20 ..........
complexity, and recognition accuracy of these algorithms
! ! Train and Toe! CLSTLK when evaluated on the alphanumeric database described in
Baseline Norm Sub CDCN
2. Recognition accuracy is somewhat different from the
Figure 1: Comparison of error rates obtained on the cen- figures reported in Fig. 1 because the version of XNIHPS
sus task with no processing, spectral subtraction, spectral used to produce these data was different. All of these al-
normalization, and the CDCN algorithm. SPHINX was gorithms are similar in function to other currently-popular
trained on the CLSTLK microphone and tested using ei- compensation strategies (e.g. 3, 9).
ther the CLSTLK microphone (solid curve) or the
The DARPA ATIS robust speech evaluation. The
PZM6FS microphone (broken curve). original CDCN algorithm described in 2 was used for the
February, 1992, ATIS-domain robust-speech evaluation.
For this evaluation, the II-XNIHPS system was trained using
Figure 1 compares the error rate obtained when the XNIHPS
the CLSTLK microphone, and tested using both the
system is trained using the DARPA standard HMD-414
CLSTLK microphone and the unidirectional Crown
closetalking microphone (CLSTLK), and tested using ei-
PCC-160 microphone (PCC160). All incoming speech in
ther the CLSTLK microphone or the omnidirectional
this evaluation was processed by the CDCN algorithm,
desktop Crown PZM-6FS microphone (PZM6FS). The
regardless of whether the testing environment was actually
census database was used, which contains simultaneous
the CLSTLK or PCC160 microphone, and the CDCN algo-
recoredings of speech from the CLSTLK and PZM6FS
rithm was not provided with explicit knowledge of the
microphones in the context of a speaker-independent
identity of the environment within which it is operating.
continuous-speech alphanumeric task with perplexity 65
2. These results demonstrate the value of the joint com- As described elsewhere in these Proceedings 10, the sys-
pensation provided by the CDCN algorithm in contrast to tem used for the official robust-speech evaluations was not
the independent compensation using either spectral sub- trained as thoroughly as the baseline system was lrained.
traction or spectral normalization. The horizontal dotted Specifically, the official evaluations were performed after
lines indicate the recognition accuracy obtained when the only a single iteration through training data that was
system is tested on the microphone with which it was processed with the CDCN algorithm, and without the
trained, with no processing. The intersection of the upper benefit of general English sentences in the training
curve with the upper horizontal line indicates that with .esabatad
CDCN compensation, XNIHPS naC recognize speech using
In Fig. 2 we show the results of an unofficial evaluation of
the PZM6FS microphone just as well when trained on the
CLSTLK microphone as when trained using the PZM6FS. the SPHINX-II system that was performed immediately
after the official evaluation was complete. The purpose of
More recently we have been attempting to develop new this second evaluation was to evaluate the improvement
algorithms which combine the computational simplicity of provided by an additional round of training with speech
SDCN with the environmental independence of CDCN. processed by CDCN, in order to be able to directly com-
One such algorithm, Blind SNR-Dependent Cepstral pare error rates on the ATIS task with CDCN with those
Normalization (BSDCN) avoids the need for environment- produced by a comparably-trained system on the same
specific training by establishing a correspondence between data, but without CDCN. As Fig. 2 shows, using the
275
CDCN algorithm causes the error rate to increase from segments with weak phonetic events, (20 percent of the
15.1% to only 20.4% as the testing microphone is changed errors were caused by cross-talk from other noise sources
from the CLSTLK to the PCC160 microphone. In contrast, in the room, and the remaining errors could not be at-
the error rate increases from 12.2% to 38.8% when one tributed to a particular cause.) Microphone arrays can, in
switches from the CLSTLK to the PCC160 microphone principle, produce directionally-sensitive gain patterns that
without CDCN. can be adjusted to produce maximal sensitivity in the
direction of the speaker and reduced sensitivity in the
,250 direction of competing sound sources. To the extent that
a
such processing could improve the effective SNR at the
input to a speech recognition system, the error rate would
G ..
be likely to be substantially decreased, because the number
30 of confusions between weak phonetic events and noise
would be sharply reduced.
~) tBeT O~t.CCP
20
Several different types of array-processing strategies have
~J~ Tesl CLSTLK been applied to automatic speech recognition. The
10 simplest approach is that of the delay-and-sum beam-
former, in which delays are inserted in each channel to
! ! compensate for differences in travel time between the
Basollno CDCN desired sound source and the various sensors (e.g.
Figure 2: Comparison of error rates obtained on the 11, 12). A second option is to use an adaptation algo-
DARPA ATIS task with no processing, spectral subtrac- rithm based on minimizing mean square energy such as the
tion, spectral normalization, and the CDCN algorithm. Frost or Gfiffiths-Jim algorithm 13. These algorithms
II-XNmPS was trained on the CLSTLK microphone in all provide the opportunity to develop nulls in the direction of
cases, and tested using either the CLSTLK microphone noise sources as well as more sharply focused beam pat-
(solid curve) or the cardiod desktop Crown PCC160 terns, but they assume that the desired signal is statistically
microphone (broken curve). independent of all sources of degradation. Consequently,
these algorithms can provide good improvement in SNR
when signal degradations are caused by additive independ-
Only two sites submitted data for the present robust speech ent noise sources, but these algorithms do not perform well
evaluation. CMU's percentage degradation in error rate in in reverberant environments when the distortion is at least
changing from the CLSTLK to the PCC160 environment, in part a delayed version of the desired speech signal
as well as the absolute error rate obtained using the 14, 15. (This problem can be avoided by only adapting
PCC160 microphone, were the better of the results from during non-speech segments 16). A third type of ap-
these two sites. proach to microphone yaarra processing is to use a cross-
correlation-based algorithm that isolates inter-sensor dif-
ferences in arrival time of the signals directly (e.g. 17).
3. MICROPHONE ARRAYS AND These algorithms are appealing because they are based on
human binaural hearing, and cross-correlation is an ef-
ACOUSTICAL PRE-PROCESSING
ficient way to identify the direction of a strong signal
source. Nevertheless, the nonlinear nature of the cross-
Despite the encouraging results that we have achieved correlation operation renders it inappropriate as a means to
using acoustical pre-processing, we believe that further im- directly process waveforms. We believe that signal
provements in recognition accuracy can be obtained in dif- processing techniques based on human binaural perception
ficult environments by combining acoustical pre- are worth pursuing, but their effectiveness for automatic
processing with other complementary types of signal speech recognition remains to be conclusively
processing. The use of microphone arrays is motivated by demonstrated.
a desire to improve the effective SNR of speech as it is
input to the recognition system. For example, the headset- Pilot evaluation of the Flanagan array. In order to ob-
mounted CLSTLK microphone produces a higher SNR tain a better understanding of the ability of array process-
than the PZM6FS microphone under normal circumstances ing to provide further improvements in recognition ac-
because it picks up a relatively small amount of additive curacy we conducted a pilot evaluation of the 23-
noise, and the incoming signal is not degraded by rever- microphone array developed by Flanagan and his col-
berated components of the original speech. leagues at AT&T Bell Laboratories. The Flanagan array,
which is described in detail inl l, 12, is a one-
To estimate the potential significance of the reduced SNR dimensional delay-and-sum beamformer which uses 32
provided by the PZM6FS microphone in the office en- microphones that are unevenly spaced in order to provide a
vironment, we manually examined all utterances in the test beamwidth that is approximately constant over the range of
set of the census task that were recognized correctly when frequencies of interest. The array uses first-order gradient
training and testing with the CLSTLK microphone but that microphones, which develop a null response in the vertical
were recognized incorrectly when training and testing plane. We wished to compare the recognition accuracy on
using the PZM6FS. We found that 54.7 percent of these the census task obtained using the Flanagan array with the
errors were caused by the confusion of silence or noise accuracy observed using the CLSTLK and PZM6FS
276
microphones. We were especially interested in determin- using the CLSTLK microphone. As expected, the worst
ing the extent to which array processing provides an im- results were obtained using the PZM6FS microphone,
provement in recognition accuracy that is complementary while the lowest error raate was obtained for speech
to the improvement in accuracy provided by acoustical recorded using the CLSTLK. More interestingly, the
pre-processing algorithms such as the CDCN algorithm. results in Fig. 3 show that both the Flanagan array and the
CDCN algorithm are effective in reducing the error rate,
.280 and that in fact the error rate at each distance obtained with
= = 3 Meters the combination of the two is very close to the error rate
obtained with the CLSTLK microphone and no acoustical
pre-processing. The complementary nature of the im-
provement of the Flanagan array and the CDCN algorithm
40
is indicated by the fact that adding CDCN to the array
improves the error rate (upper panel of Fig. 3), and that
20 converting to the array even when CDCN is already
employed also improves performance (lower panel).
0 ! | ! m !
PZM6FS Array Array CLSTLKCLSTLK
NCDC.4 +CDCN
Microphone Type 4. PHYSIOLOGICALLY-MOTIVATED
= r"2"'=
FRONT ENDS AND
O8,99,.
= = 3 Meters ACOUSTICAL PRE-PROCESSING
In recent years there has also been an increased interest'in
O4 the use of peripheral signal processing schemes that are
motivated by human auditory physiology and perception,
20 and a number of such schemes have been proposed (e.g.
18, ,91 20, 21). Recent evaluations indicate that with
"clean" speech, such approaches tend to provide recog-
0 I I I I t
SF6MZP PZM6FS Array CLSTLK CLSTLK nition accuracy that is comparable to that obtained with
+CDCN +CDCN NCDC+ conventional LPC-based or DFT-based signal processing
Microphone Type
schemes, but that these auditory models can provide
Figure 3: Comparison of recognition accuracy obtained greater robustness with respect to enviromental changes
on a portion of the census task using the omnidirectional when the quality of the incoming speech (or the extent to
Crown PZM-6FS, the 23-microphone array developed by which it resembles speech used in training the system)
Flanagan, and the Senneheiser microphone, each with decreases 22, 23. Despite the apparent utility of such
and without CDCN. Data were obtained from simul- processing schemes, no one has a deep-level understanding
taneous recordings using the three microphones at dis- of why they work as well as they do, and in fact different
tances of 1 and 3 meters (for the PZM-6FS and the array). researchers choose to emphasize rather different aspects of
the peripheral auditory system's response to sound in their
work. Most auditory models include a set of linear
41 utterances from the census database were obtained from bandpass filters with bandwidth that increases nonlinearly
each of five male speakers in a sparsely-furnished with center frequency, a nonlinear rectification stage that
laboratory at the Rutgers CAIP Center with hard walls and frequently includes short-term adaptation and lateral sup-
floors. The reverberation time of this room was informally pression, and, in some cases, a more central display based
estimated to be between 500 and 750 ms. Simultaneous on short-term temporal information. We estimate that the
recordings were made of each utterance using three number of arithmetic operations of some of the currently-
microphones: the Sennheiser HMD-414 (CLSTLK) popular auditory models ranges from 53 to 600 times the
microphone, the Crown PZM6FS, and the Flanagan array number of operations required for the LPC-based process-
with input lowpass-filtered at 8 kHz. Recordings were ing used in .II-XNmPS
made with the speaker seated at distances of ,1 2, and 3
Pilot evalution of the Seneff auditory model. We
meters from the PZM6FS and Flanagan array
recently completed a series of pilot evaluations using an
microphones, wearing the CLSTLK microphone in the
implementation of the Seneff auditory model 21 on the
usual fashion at all times.
census databse. Since almost all evaluations of
Figure 3 summarizes the error rates obtained from these physiologically-motivated front ends to date have been
speech samples at two distances, 1 and 3 meters, with and performed using artificaUy-added white Gaussian noise,
without the CDCN algorithm applied to the output of the we have been interested in the extent to which auditory
microphone army. Error rates using the CLSTLK models can provide useful improvements in recognition
microphone differed somewhat for the two distances be- accuracy for speech that has been degraded by reverbera-
cause different speech samples were obtained at each dis- tion or other types of linear filtering. As in the case of
tance and because the sample size is small. The XNIHPS microphone arrays, we are also especially interested in
system had been previously trained on speech obtained determining the extent to which improvements in robust-
277
ness provided by auditory modelling complement those previous studies. The results in the lower panel of Fig. 4,
that we already enjoy by the use of acoustical pre- demonstrate that the mean rate and GSD outputs of the
processing algorithms such as CDCN. Seneff model provide lower error rates than conventional
LPC cepstra when the system is trained using the CLSTLK
We compared error rates obtained using the standard 12
microphone and tested using the PZM6FS. Nevertheless,
LPC-based cepstral coefficents normally input to the
the level of performance achieved by the present im-
XNIHPS system, with those obtained using an implemen-
plementation of the auditory model is not as good as that
tation of the 40-channel mean-rate output of the Seneff
achieved by conventional LPC cepstra combined with the
model 21, and with the 40-channel outputs of Seneff's
CDCN algorithm on the same data (Fig. .)1 Furthermore,
Generalized Synchrony Detectors (GSDs). The system
the combination of conventional LPC-based processing
was evaluated using the original testing database from the
and the CDCN algorithm produced performance that
census task with the CLSTLK and PZM6FS microphones,
equaled or bettered the best performance obtained with the
and also with white Ganssian noise artificially added at
auditory model for each test condition. Because the
sig.nal-to-noise ratios of +10, +20, and +30 dB, measured
auditory model is nonlinear and not easy to port from one
using the global SNR method described in 19.
site to another, these comparisons should all be regarded as
preliminary. It is quite possible that performance using the
100 Z~
auditory model could further improve if greater attention
... 80 J." "~ LPC were paid to tuning it to more closely match the charac-
I~ ~ , ~"'~' LPC+CDCN teristics of .XNIHPS
60 ,me Mean Rate
%% : : GSD We also attempted to determine the extent to which a com-
4o
bination of auditory processing and the CDCN algorithm
2O could provide greater recognition accuracy than either
processing scheme used in isolation. In these experiments
0 lO'dB 20'dB 30'dB CI;an we combined the effects of CDCN and auditory processing
SNR by resynthesizing the speech waveform from cepstral coef-
ficients that were produced by the original LPC front end
,~. 100
and then modified by the CDCN algorithm. The resyn-
@ os thesized speech, which was totally intelligible, was then
passed through the Seneff auditory model in the usual
fashion. Unfortunately, it was found that this particular
4O k- ~ CPL combination of CDCN and the auditory model did not im-
prove the recognition error etaar beyond the level achieved
20 : = Mean Rate by CDCN alone. A subsequent error analysis revealed that
- - GSD
this concatenation of cepstral processing and the CDCN
10 ' dB 20 ' dB 30'dB Clean ' algorithm, followed by resynthesis and processing by the
SNR original XNIHPS front end, degraded the error rates even in
Figure 4: Pilot data comparing error rates obtained on the the absence of the auditory processing, although analysis
census task using the conventional LPC-based processing and resynthesis without the CDCN algorithm did not
of SPHINX with results obtained using the mean rate and produce much degradation. This indicates that useful in-
synchrony outputs of the Seneff auditory model. XNIHPS formation for speech recognition is lost when the resyn-
was trained on the CLSTLK microphone in all cases, and thesis process is performed after the CDCN algorithm is
run. Hence we regard this experiment as inconclusive, and
tested using either the CLSTLK microphone (upper panel)
we intend to explore other types of combinations of acous-
or the Crown PZM6FS microphone (lower panel). White
tical pre-processing with auditory modelling in the future.
noise was artificially added to the speech signals and data
are plotted as a function of global SNR.
5. SUMMARY AND CONCLUSIONS
Figure 4 summarizes the results of these comparisons, with
error rate plotted as a function of SNR using each of the In this paper we describe our current research in acoustical
three peripheral signal processing schemes. The upper pre-processing for robust speech recognition, as well as
panel describes recognition error rates obtained with the our first attempts to integrate pre-processing with other
system both trained and tested using the CLSTLK approaches to robust speech recognition. The CDCN algo-
microphone, and the lower panel describes error rates ob- rithm was also applied to the ATIS task for the first time,
tained with the system trained with the CLSTLK and provided the best recognition scores for speech col-
microphone but tested with the PZM6FS microphone. lected using the unidirectional desktop PCC160
When the system is trained and tested using the CLSTLK microphone. We demonstrated that the CDCN algorithm
microphone, best performance is obtained using conven- and the Flanagan delay-and-sum microphone array can
tional LPC-based signal processing for "clean" speech. As provide complementary benefits to speech recognition in
the SNR is decreased, however, error rates obtained using reverberant environments. We also found that the Seneff
either the mean rate or GSD outputs of the Seneff model auditory model improves recognition accuracy of the CMU
degrade more gradually confirming similar findings from speech system in reverberant as well as noisy environ-
278
ments, but preliminary efforts to combine the auditory .01 Ward, W., Issar, S., Huang, X., Hon, H.-W., Hwang,
model with the CDCN algorithm were inconclusive. M.-Y., Young, S., Matessa, M., Liu, F.-H., Stem, R.,
"Speech Understanding in Open Tasks", Proc. of the
DARPA Speech and Natural Language Workshop,
ACKNOWLEDGMENTS February 1992.
.11 Flanagan, J. L., Johnston, J. D., Zahn, R., and Elko, G.W.,
This research is sponsored by the Defense Advanced Research
"Computer-steered Microphone Arrays for Sound
Projects Agency, DoD, through APRA Order 7239, and monitored
Transduction in Large Rooms", The Journal of the
by the Space and Naval Warfare Systems Command under con-
Acoustical Society of America, Vol. 78, Nov. 1985, pp.
tract N00039-91-C-0158. Views and conclusions contained in
.8151-8051
this document are those of the authors and should not be inter-
preted as representing official policies, either expressed or im- .21 Flanagan, J. L., Berkeley, D. A., Elko, G. W., West, J. ,.E
plied, of the Defense Advanced Research Projects Agency or of and Sondhi, M. M., "Autodirective microphone sys-
the United States Government. We thank Hsiao-Wuen Hon, terns", Acustica, Vol. 73, February 1991, pp. 58-71.
Xuedong Huang, Kai-Fu Lee, Raj Reddy, Eric Thayer, Bob
.31 Widrow, B., and Stearns, S. D., Adaptive Signal
Weide, and the rest of the speech group for their contributions to
Processing, Prentice-Hall, Englewood Cliffs, NJ, 1985.
this work. We also thank Jim Flanagan, Joe French, and
A. C. Surendran for their assistance in obtaining the experimental .41 Peterson, P. M., "Adaptive Array Processing for Multiple
data using the array microphone, and Stephanie Seneff for Microphone Hearing Aids". RLE TR No. 541, Res. Lab.
providing source code for her auditory model. The graduate of Electronics, MIT, Cambridge, MA
studies of Tom Sullivan and Yoshlaki Ohshima have been sup-
ported by Motorola and IBM Japan, respectively. .51 Alvarado, V. M., Silverman, H. F., "Experimental
Results Showing the Effects of Optimal Spacing Between
Elements of a Linear Microphone Array", ICASSP-90,
REFERENCES April 1990, pp. 837-840.
.61 Van Compernolle, D., "Switching Adaptive Filters for
.1 Juang, B. H., "Speech Recognition in Adverse Envoron- Enhancing Noisy and Reverberant Speech from
ments", Comp. Speech and Lang., Vol. 5, 1991, pp. Microphone Array Recordings", IEEE International Con-
275 -294. ference on Acoustics, Speech, and Signal Processing,
April 1990, pp. 833-836.
2. Acero, A. and Stern, R. M., "Environmental Robusmess
in Automatic Speech Recognition", ICASSP-90, April .71 Lyon, R. F., "A Computational Model of Binaural
,0991 pp. 849-852. Localization and Separation", IEEE International Con-
ference on Acoustics, Speech, and Signal Processing,
3. Erell, A. and Weintraub, M., "Estimation Using Log-
,3891 pp. 1148-1151.
Spectral-Distance Criterion for Noise-Robust Speech
Recognition", ICASSP-90, April 1990, pp. 853-856. .81 Cohen, J. R., "Application of an Auditory Model to
Speech Recognition", The Journal of the Acoustical
4. Acero, A. and Stem, R. M., "Robust Speech Recognition
Society of America, Vol. ,58 No. ,6 June 1989, pp.
by Normalization of the Acoustic Space", ICASSP-91,
2623 -2629.
May 1991, pp. 893-896.
.91 Ghitza, O., "Auditory Nerve Representation as a Front-
5. Liu, F.-H., Acero, A., and Stern, R. M., "Efficient Joint
End for Speech Recognition in a Noisy Environment",
Compensation of Speech for the Effects of Additive Noise
.paroC Speech and Lang., Vol. ,1 1986, pp. 109-130.
and Linear Filtering", IEEE International Conference on
Acoustics, Speech, and Signal Processing, March 1992. 20. Lyon, R. F., "A Computational Model of Filtering,
Detection, and Compression in the Cochlea", IEEE Inter-
.6 Stockham, T. G., Cannon, T. M,. and lngebretsen, R. B.,
national Conference on Acoustics, Speech, and Signal
"Blind Deconvolution Through Digital Signal Process-
Processing, May 1982, pp. 1282-1285.
ing", Proc. IEEE, Vol. ,36 1975, pp. 678-692.
21. Seneff, S., "A Joint SynchronyMean-Rate Model of
7. Boll, .S F. , "Suppression of Acoustic Noise in Speech
Auditory Speech Processing", Journal of Phonetics, Vol.
Using Spectral Subtraction", ASSP, Vol. 27, 1979, pp.
,61 No. ,1 January 1988, pp. 55-76.
.021-311
22. Hunt, M., "A Comparison of Several Acoustic Represen-
8. Sakoe, H., Chiba, S., "Dynamic Programming Algorithm
tations for Speech Recognition with Degraded and Un-
Optimization for Spoken Word Recognition", IEEE
degraded Speech", IEEE International Conference on
Transactions on Acoustics, Speech, and Signal
Acoustics, Speech, and Signal Processing, May 1989.
Processing, Vol. 26, 1978, pp. 43-49.
23. Meng, H., and Zue, V. W., "A Comparative Study of
9. Hermansky, H., Morgan, N., Bayya, A., Kohn, P., "Com-
Acoustic Representations of Speech for Vowel Classifica-
pensation for the Effect of the Communication Channel in
tion Using Multi-Layer Perceptrons", Int. Conf. on
Auditory-Like Analysis of Speech (RASTA-PLP)", Proc.
Spoken Lang. Processing, November 1990, pp.
of the Second European Confi on Speech Comm. and
.6501-3501
Tech., September 1991.
279